An Anthropic researcher just quit and said the quiet part out loud

Jacob Coxon spent three years doing pretraining research at OpenAI and Anthropic before resigning this week, telling employees and then the public that both companies are racing toward self-improving systems while gambling with outcomes they cannot control. He is 27, and his job was literally to help scale the models he now says nobody can safely scale further. What makes this land differently from the usual AI doom cycle is the source: not a policy adviser or an outside critic, but someone who was inside the pretraining pipeline at two of the three labs setting the pace. Coxon is calling for pacing agreements between labs, a voluntary slowdown that has approximately zero chance of being adopted while OpenAI, Anthropic and Google remain locked in a capital arms race measured in tens of billions. Anthropic has said nothing to rebut the substance of his claims, only that it takes safety seriously, which is the same thing it said before he left.